Scholarlink

<p>Medical image segmentation underpins diagnosis, treatment planning, and disease monitoring, yet the deep learning models that achieve the highest accuracy are typically too large and computationally demanding for the clinical workstations, mobile diagnostic devices, and point-of-care systems on which they must ultimately run. Closing this gap between research-grade accuracy and clinical hardware constraints has become a central challenge for the field. In this survey, we present a unified, efficiency-focused review of deep learning methods for medical image segmentation, organising over 150 works (2015–2026) within a single framework that spans convolutional, Transformer, and Mamba/state-space architectures together with the compression techniques that render them deployable. To the best of our knowledge, no prior survey jointly covers all three architectural families alongside model compression and controlled empirical benchmarking. We structure the field through a four-pillar taxonomy: six families of efficient architectural strategy; model compression and deployment (knowledge distillation, pruning, quantization, and hardware-aware neural architecture search, with tiered guidelines for edge, workstation, and cloud); domain challenges, clinical loss formulations, and data-efficient learning; and a standardised evaluation framework of segmentation metrics, benchmark datasets, and statistical validation. To ground the review in direct evidence, we further conduct a controlled cross-modality benchmark in which representative architectures from all six families are retrained from scratch under a single protocol across dermoscopic and endoscopic polyp datasets, with boundary-quality and deployment-cost profiling. The experiments expose two trade-offs that single-dataset, overlap-only evaluation conceals: compact encoder–decoder models remain competitive with far larger architectures on well-standardised binary tasks, while channel-attention designs that excel on large datasets can degrade sharply when training data is scarce a data-efficiency failure invisible to parameter and computation budgets. These observations motivate tiered deployment guidelines and five open problems: boundary-preserving compression, domain generalisation, clinical interpretability, data-efficient learning for rare pathologies, and real-time volumetric inference. We additionally release a GitHub project page compiling key resources for efficient deep learning in medical image segmentation: <a href="https://github.com/razanharith/efficient-medseg">https://github.com/razanharith/efficient-medseg</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Deep Learning for Medical Image Segmentation: A Unified Cross-Paradigm Survey and Controlled Deployment Benchmark of CNN, Transformer, and Mamba Architectures

  • 期刊论文

  • 发表日期：

    2026年8月5日

查看全文
  • Razan Alharith,
  • Mugahed A. Al-antari,
  • Kaleem Ullah Qasim,
  • Zaid Al-Huda,
  • Yeong Hyeon Gu

摘要

Medical image segmentation underpins diagnosis, treatment planning, and disease monitoring, yet the deep learning models that achieve the highest accuracy are typically too large and computationally demanding for the clinical workstations, mobile diagnostic devices, and point-of-care systems on which they must ultimately run. Closing this gap between research-grade accuracy and clinical hardware constraints has become a central challenge for the field. In this survey, we present a unified, efficiency-focused review of deep learning methods for medical image segmentation, organising over 150 works (2015–2026) within a single framework that spans convolutional, Transformer, and Mamba/state-space architectures together with the compression techniques that render them deployable. To the best of our knowledge, no prior survey jointly covers all three architectural families alongside model compression and controlled empirical benchmarking. We structure the field through a four-pillar taxonomy: six families of efficient architectural strategy; model compression and deployment (knowledge distillation, pruning, quantization, and hardware-aware neural architecture search, with tiered guidelines for edge, workstation, and cloud); domain challenges, clinical loss formulations, and data-efficient learning; and a standardised evaluation framework of segmentation metrics, benchmark datasets, and statistical validation. To ground the review in direct evidence, we further conduct a controlled cross-modality benchmark in which representative architectures from all six families are retrained from scratch under a single protocol across dermoscopic and endoscopic polyp datasets, with boundary-quality and deployment-cost profiling. The experiments expose two trade-offs that single-dataset, overlap-only evaluation conceals: compact encoder–decoder models remain competitive with far larger architectures on well-standardised binary tasks, while channel-attention designs that excel on large datasets can degrade sharply when training data is scarce a data-efficiency failure invisible to parameter and computation budgets. These observations motivate tiered deployment guidelines and five open problems: boundary-preserving compression, domain generalisation, clinical interpretability, data-efficient learning for rare pathologies, and real-time volumetric inference. We additionally release a GitHub project page compiling key resources for efficient deep learning in medical image segmentation: https://github.com/razanharith/efficient-medseg.